**The most surprising finding is that the big "surface effect" on serving is confounded by who plays on each surface.** These numbers come from the surface-level summary of `raw.raw_kaggle` that I already pulled (rows are player-per-match). - **Aces differ sharply by surface.** Average aces per player-match are about 3.4 on clay, 5.9 on hard, 7.6 on grass and 7.9 on carpet. Grass and carpet players hit more than twice as many aces as clay players. Average service points follow the same pattern (84 on grass vs 73 on clay), so grass matches may also be longer, which would inflate ace counts further. - **Double faults barely move.** They sit in a narrow band of about 2.8 to 3.2 per match across all four surfaces. The surface seems to change how often a serve can't be returned, but not how often players miss the serve. - **The player pool differs by surface.** The average ranking of the player in the row (lower is better) is about 151 on grass, 311 on hard, 361 on clay and 517 on carpet. The median is 109 on grass against 288 on clay. Grass has only 9,479 rows, about 4% of the data, and they come mostly from top-ranked players. Clay has 107,650 rows and hard has 117,571. - **The high-rank skew is not just a grass quirk.** On every surface, the average rank of the player in the row is better than their opponent's (for example 311 vs 379 on hard). So the ace gap cannot be read as a pure "grass is fast, clay is slow" effect. It is partly a selection effect, because grass matches in this data are played by a much stronger group of players, and stronger players serve more aces. A surface-based scouting or betting model should control for rank and player before trusting the surface coefficients. **Caveat:** I have not yet run the rank-controlled comparison, so I can't say how much of the ace gap survives it. Separately, 86,793 of the 237,205 rows (about 37%) have no serve statistics (`TP`, `Aces`, `DFs`, `SP`, `1SP`, `2SP` and `vA` are all null together). Surface averages for those measures therefore rest on a non-random subset of matches. Another 28 rows have a blank surface.